Dual-scan self-learning denoising for application in ultralow-field MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Ultralow field (ULF) magnetic resonance imaging (MRI) presents unique challenges in image quality due to inherent noise.
• Effective denoising is crucial for accurate diagnosis and quantitative analysis in ULF MRI applications.

Purpose of the Study:

• To develop and validate a self-learning method for denoising MR images specifically for ULF applications.
• To improve the performance of denoising algorithms beyond traditional methods in ULF environments.

Main Methods:

• A self-learning neural network approach is proposed, utilizing dual-acquisition MRI data as training pairs.
• The method is based on the Noise2Noise framework, incorporating enhanced data augmentation and an integrated learning strategy.
• The model is trained and evaluated on both synthetic and real ULF MRI datasets.

Main Results:

• The proposed self-learning model demonstrates superior denoising performance compared to the traditional Noise2Noise method, both subjectively and objectively.
• Magnitude images from ULF MRI are effectively denoised, outperforming several state-of-the-art methods.
• The method shows improved results for phase images and quantitative imaging applications, attributed to its self-learning framework.

Conclusions:

• The developed self-learning model significantly enhances magnitude image denoising for ULF MRI using real-world data.
• The method's effectiveness extends to phase and quantitative imaging, outperforming existing denoisers.
• This self-learning framework offers a robust solution for improving ULF MRI image quality and diagnostic utility.